What Is Field of View in Egocentric Video?

Two datasets can both be called egocentric and still differ in how much of the task each frame holds. Field of view sets the crop: how much scene context surrounds the hands, and how far to the side an object can sit before it leaves the frame. A buyer who knows only resolution and frame rate cannot judge whether the views will contain what the model needs to see, so the angle, the axis it is quoted on and the camera it applies to are worth asking for in writing.

The head-mounted global-shutter 4K camera is stated at a 120° field of view, shown as 120° HFOV on the capture rig diagram. The two wrist cameras are described as having a narrow field of view, and the depth and stereo views have no stated angle. Each camera has its own intrinsics (fx, fy, cx, cy and distortion terms) solved before and after capture, and all streams are hardware-triggered onto one clock. Any other angle, or an angle after rectification, is Not stated and belongs in the spec of a custom or on-demand collection.

Field of view, or FOV, is the angular extent of the scene a camera can see, measured in degrees. It can be quoted horizontally, vertically or diagonally, and a figure is only meaningful once it says which. The site abbreviates the horizontal form as HFOV.

In egocentric video the camera is worn by the person doing the task, so it moves with their head and shares their field of view. The angle therefore decides how much of the hands, the workpiece and the surrounding scene stays inside the frame at any moment, and what is cropped out entirely.

A head camera exists to carry scene context: the environment, the target object and the approach. A wider view keeps more of that context in frame, so a policy can see where the hands are heading and what surrounds them. A narrower view keeps less context but gives more of the frame to whatever is directly ahead.

The site also states the limit of any head view: as a hand closes on an object, the head camera’s line of sight to the contact point is frequently blocked by the hand and arm. The site treats that occlusion as a structural weak point of the head view and answers it with wrist cameras rather than with a different angle. The wrist cameras have a narrow field of view that keeps the grasp point in frame but loses the scene outside it.

Field of view is therefore a coverage trade-off, not a quality score. The site does not rank angles. It pairs a wide head view with narrow wrist views so that each covers what the other loses.

OpenCV’s camera model describes a camera by an intrinsic matrix of focal lengths fx and fy, expressed in pixel units, and a principal point (cx, cy) that is usually close to the image centre. OpenCV also gives the conversion between angle and lens: the physical focal length can be deduced from the angular field of view, as the sensor size divided by twice the tangent of half the angle.

The same documentation lists calibrationMatrixValues, a function that computes useful camera characteristics, including fovx and fovy, from the intrinsic matrix, the image size and the sensor aperture dimensions. So once a camera is calibrated, its field of view can be derived from the intrinsics rather than needing a separate measurement. OpenCV adds that if an image is scaled, all of these intrinsic parameters must be scaled by the same factor, so focal lengths in pixels are only comparable between images of the same resolution.

OpenCV states that real lenses usually have some distortion, mostly radial with slight tangential, so its model extends the pinhole projection with distortion coefficients. Those coefficients do not depend on the scene viewed and stay the same regardless of image resolution. calibrationMatrixValues takes no distortion coefficients as input, so the angle it returns follows the ideal pinhole model rather than the distorted image.

The site’s lens distortion entry notes that radial distortion is largest toward the frame edge. For a wide view, how far to trust edge geometry depends on how the distortion was modelled. The site lists per-camera distortion terms (k1, k2, p1, p2, k3) in each episode’s calibration group but does not state their values.

One angle is stated. The head-mounted global-shutter 4K camera has a 120° field of view, labelled 120° HFOV on the capture rig diagram, and the capture summary describes the same camera as 4K at 60 fps with a 120° lens. Because the label is HFOV, the stated figure is the horizontal angle.

Everything else is not stated. That includes the vertical and diagonal angles, the lens model, the field of view of the wrist cameras, the depth sensor and the stereo rig, and the angle after any undistortion or rectification. This page does not fill those in.

The site describes the head view as wide and the wrist view as narrow, in words rather than angles. The head camera gives the wearer’s wide egocentric view of the environment, target object and approach. Each of the two wrist cameras gives a close view of that hand’s grasp point and loses scene context outside its narrow field of view.

For stereoscopic capture, the site describes a head-mounted stereo rig at a human interpupillary baseline, calibrated per session. Depth from disparity can only be recovered for scene points that both cameras see, so the overlap between the two views matters as well as each camera’s own angle. The site does not state that overlap, the stereo lens angle, or whether the stereo rig uses the same 120° head camera described above.

Custom and on-demand collections run against a written spec and are QA-checked against it. If a field of view matters to your model, write the angle and the axis it applies to into the spec, along with the views it covers (head, wrist or stereo) and whether you need the angle after rectification. Whether a requested angle is available is confirmed against the spec; this page does not promise one.

On timing, the site states a median of 48 hours from a signed spec to the first validated batch where coverage is already deep for the task and environment mix, and three to four weeks where coverage is thin. Those timelines depend on coverage depth for the task and environment mix.

What is field of view in egocentric video?

It is the angular extent of the scene the worn camera can see, in degrees. It decides how much of the hands, workspace and surroundings stays in frame.

What field of view does Firsthand state for its head camera?

120°, shown as HFOV (horizontal field of view) on the capture rig diagram, for a head-mounted global-shutter 4K camera.

Does Firsthand state the field of view of its wrist, depth or stereo cameras?

Not stated. The wrist cameras are described as a narrow field of view and the stereo rig by its human interpupillary baseline, but no angles are given for either, or for the depth sensor.

Is a wider field of view always better for robot learning?

The site does not rank angles. It pairs a wide head view for scene context with narrow wrist views for contact detail, so each covers what the other loses.

How can field of view be worked out from camera calibration?

OpenCV documents that focal length and angular field of view can be converted into each other, and its calibrationMatrixValues function returns fovx and fovy from the intrinsic matrix, image size and aperture dimensions. The site lists per-camera intrinsics but does not state their values.

Can I request a specific field of view in a custom collection?

Write the angle, the axis and the camera into the spec for a custom or on-demand collection. Availability is confirmed against the spec, so request a custom dataset and state the requirement up front.